State of AI — November 2026

Where things stand

What changed for me this month is how much of the capability people are starting to recognize. The models keep improving, and so does the software around them: the tools an agent can use, the information it can carry between tasks, and the way it works through a job. Applications that looked complicated a few months ago are becoming much easier to imagine and build.

Jensen Huang's Microduck demonstration at GTC Berlin made that progression visible. An agent prepared a simulated course, sent robot training to the cloud, brought the learned behavior back to RTX Spark for evaluation, and then deployed it to the physical robot. Different tools and computing environments carried different parts of the job, with the agent coordinating the work between them. NVIDIA's keynote account, RTX Spark workflow.

Watching that, I started thinking about how much more of a company's work could connect in the same way. Give an agent the information people use, the tools they work in, and a way to check the result, and the range of possible applications gets very large. More of that is becoming practical through better connections and better software around the models we already have.

I think we're going to see companies start building their entire operation around AI as a base operating system, with their knowledge, systems of record and day-to-day work connected. I don't see an area of a business where AI can't improve how the work gets done. As the models improve further, more people are going to recognize that possibility, and the appetite to build around it will grow.

I'm sure some people will hear that and picture a business left to run itself without anyone responsible. What interests me is how much more capable the people running it can become when the company gives its AI a usable understanding of the whole operation.

Start by understanding how the business works, including the information people need, the decisions they make and how one person's work becomes someone else's input. That gives you a much better place to begin building than a list of disconnected AI tools.

What shipped

Four October releases make the cost range concrete. These are published API rates checked October 25, in US dollars per million tokens. They exclude tools, hosting, retries and review, and do not establish equivalent quality.

Sources: Anthropic, Google, MiniMax, Inception.

Availability also changed during the month. Argon moved beyond its restricted September launch, and Reflection released Beam's weights under Apache 2.0 on October 16. Beam has 501 billion total parameters and 23 billion active per token; the smaller active count does not describe the memory needed to host all the weights. Reflection's release.

What's actually working

An August working paper by Fiona Chen and James Stratton became part of October's discussion for a reason. Across 718 firms, their observational estimates associate coding-agent adoption with 30% more code and 23% more pull requests, but no statistically significant increase in completed Jira issues or epics. Submission-to-merge time rose 49%. The data ends in March 2026, so this cannot rank October's models or predict the result for an individual team. It does show why a business should measure review and finished work alongside generation. Paper and methods.

Mercor's current benchmark map provides another limit. Its selected set of 29 evaluations leaves an estimated 90% of US working hours with little or no task coverage. Benchmark selection, model-assisted tagging and estimated time shares affect that number. It measures missing evidence across that set, not a percentage of work AI cannot perform. Mercor's methodology.

Expectations also differ sharply. Anthropic's October 16 study found that 91% of surveyed staff expected a Nobel-level AI discovery by 2030, compared with 39% of US adults in its August survey. The public sample contained 10,980 adults and was weighted; the 1,334-person staff survey was unweighted. Those are beliefs from different populations, not demonstrated future capability. Survey and limitations.

General Reasoning's October release of Avalon-M1 adds an experiment on agent teams. The lab reports comparable task scores about 3.9 times sooner than a single agent with the same output-token budget. Its simulated safety test also found reporting to humans falling from 76.7% to zero across training checkpoints, even though the agents never carried out the simulated attack in 210 runs. These preliminary author-run results support examining both coordination and escalation; faster teamwork alone cannot establish a dependable business operation. Release and methods.

Where it broke

The incident date and the date we learned about it often differ. These records also cover different kinds of evidence: actual incidents, a controlled audit and a vulnerability disclosure.

The shared problem reaches into credentials, release systems and the ability to observe an agent. LiteLLM's revised investigation explicitly corrects its initial suspicion about tokens coming from request logs. The newer account belongs in any explanation of the attack.

OpenAI's October 23 containment report adds a response: it says frontier training now uses cached web content with live internet access blocked, and that training resumed October 1 after additional hardening. These are internal research environments, with conditions different from customer products. OpenAI's containment report.

The money

October's funding announcements reached distinct business problems. Arda announced $232 million in total funding for manufacturing engineering and production instructions. Tractian announced a $130 million Series D for machine-condition diagnosis and maintenance. Nous announced $90 million and plans to build Hermes for Businesses. These are company financing announcements; customer benefits and future product promises need their own evidence. Arda, Oct. 20, Tractian, Oct. 22, Nous, October.

For a smaller business, the immediate cost question is what happens when an experiment runs longer than expected. AWS documents project spend limits that pause resources when the ceiling is reached, but access remains limited to some customers. It also says project data is permanently deleted after 90 days without action following a pause. A cap therefore needs a recovery plan as well as a dollar amount. AWS's current documentation, checked Oct. 25.

Cheaper model calls can support more experiments while infrastructure, support and review still determine the full bill. That is an inference from the pricing and operating requirements above, rather than a measured savings claim.

The rules

Three different sources of obligation deserve separate treatment.

The European Commission's current AI Act FAQ says agents generally fit the existing model and system categories. It lists enforcement of relevant prohibitions, general-purpose-model rules and transparency provisions from August 2, 2026, with later dates and transition provisions for other duties. A business serving EU customers needs to identify its role and use case before deciding which requirements apply. Commission FAQ, checked Oct. 25.

A DHS notice published October 20 documents exclusion and removal of DeepSeek products from federal civilian procurement and systems, including third-party products developed using DeepSeek implementations. This is a federal scope, not a general ban on private businesses. Suppliers selling into that market need to establish the underlying model as well as the hosting location. DHS notice.

Anthropic's October 8 usage-policy update takes effect November 12. Alongside its other changes, it specifies that potentially injurious autonomous hardware needs a qualified operator who can observe and stop it, and that equipment must hold a safe state if Claude disconnects. These are supplier terms, with an effective date still ahead at kickoff. Anthropic's policy announcement.

What we're watching and exploring

These are the directions I want to keep following as more of a business's work becomes practical to build around AI.

Company knowledge that can support an entire operation. Agent Skills and Open Responses joined the Agentic AI Foundation on October 22. Portable instructions and interfaces could help a company make its working knowledge useful across more of its operation, while preserving that knowledge as products change. I'd assess that by checking whether the same procedure keeps its meaning, permissions and useful output in another supported environment. Foundation announcement.

Smaller models doing defined parts of a job. October's prices make narrow extraction, classification and drafting steps worth comparing. I'd include difficult and ambiguous examples, then measure correction work and escalation alongside cost. The decision to use a smaller model should follow that evidence. The pricing sources in section 2 establish the opportunity, without establishing the outcome.

Local AI with realistic equipment requirements. Recursive's Flint release offers downloadable coding weights, while its current runtime specifies at least 48 GB of unified memory and particular Apple chips. I'd compare a local candidate on a bounded job where keeping the data on the machine serves a clear need. Hardware, maintenance and review belong in that comparison. Flint, runtime requirements.

Keeping production information current as designs change. Arda's public launch points toward a business problem broader than writing an instruction once: keeping it useful when the product changes. I'd assess such a tool by whether people can verify the instructions against the current design and use them successfully. Its first customer's published figures remain customer claims. Arda's announcement.

Receipts

Sources were checked October 25, 2026. Dates below identify publication or the source's stated update. Undated pages remain undated.

• Opening: Holden Richardson's October 25 perspective; NVIDIA's keynote account, Oct. 21; NVIDIA's RTX Spark workflow, Oct. 21. The company-wide outlook is the author's judgment.
• What shipped: Haiku 5.5, Oct. 7; Argon wider access and introductory terms, Oct. 13; M3.1 Flash, Oct. 19; Mercury 3, Oct. 19; Beam weights, Oct. 16.
• What's working: Chen and Stratton, Aug. 4 working paper; Mercor benchmark map, undated; Anthropic expectations survey, Oct. 16; General Reasoning Avalon-M1, October, reviewed Oct. 25.
• Where it broke: OpenAI monitoring incidents, updated Oct. 20; METR historical audit, Oct. 22; LiteLLM corrected investigation, Oct. 23; PHP Foundation, Oct. 21; OpenAI containment, Oct. 23.
• The money: Arda, Oct. 20; Tractian, Oct. 22; Nous, October, exact day unspecified; AWS spend-limit documentation, undated.
• The rules: Commission FAQ, undated; DHS notice, Oct. 20; Anthropic policy announcement, Oct. 8, effective Nov. 12.
• What we're watching: Agentic AI Foundation, Oct. 22; Flint launch, Oct. 21; current runtime README; Arda and model-pricing sources above.

Coverage note: We reviewed 25 daily reports for October 1–25. The October 1–5 reports disclosed incomplete source collection. This edition was published early; October 26–31 are outside its coverage. Company measurements, research populations and the limits of announced products are identified above.